[Yigtwxx]

What this changes

• A short rate limit is no longer permanent. A 429 whose Retry-After is at most 120 s (60 s when the header is absent) switches the judge to the local model only until that time has passed. The first request after that probes the provider. If it answers, the judge is back on it, and a one-time notice says so ("Groq answering again — local ollama qwen3.5:9b stood in for N requests"). Only one thread probes; the others keep using the local model until the probe has answered.
• Everything else is sticky, as before: a long 429 (daily limit), 413, 401/403, 5xx, a timeout, or no answer.
• Honest attribution. Each Completion now carries the client that answered it, so an opinion in flight across a switch or a return is named for the model that actually gave it (rule 6). Before, it took whatever the judge was routed to when the opinion was read.
• A dead local model stays dead. It is never re-announced as "judging", it is not warmed again, and the return notice says it failed instead of "stood in for 0 requests".
• Retry-After: 0 reads "~1s". The waiters of a re-switch are woken on the right event.
• Docs: spec docs/superpowers/specs/2026-10-05-judge-cooldown-design.md (amends evidence-search spec §11), OPEN-ITEMS §22 (with one follow-up left open on purpose), CHANGELOG, README.

Why

The user asked whether the judge could go back to Groq once its minute limit resets, instead of staying on qwen3.5:9b for the rest of the run. Groq's free tier for gpt-oss-120b is 8K tokens per minute, about one full batch. They chose option A, a cooldown with return. Both goals of the original §11 decision hold: nobody waits, and the provider is asked again only once the time it named has passed.

An independent review found three important bugs in the first version: a dead local model re-announced as judging, opinions misattributed in both directions, and a lost wake-up on _ready. It also found four nits. A re-review confirmed the fixes and was clean.

Checklist

• ruff check and ruff format --check are clean (and mypy --strict)
• pytest passes (full suite; the threading tests ran 25 times without a flake)
• Type annotations on every new function signature
• No network access in unit tests, no real sleeps (injected clock)
• If a new reported state was added, spec section 15 is updated too (no new verification state; new judge notices only)
• This change does not make the tool sound more confident than its evidence

🤖 Generated with Claude Code

https://claude.ai/code/session_01K46EEsxMvuAuM8wjvPNWLM

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
proofpath	Ignored		Oct 5, 2026 12:54pm UTC